The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from GitHub CI logs and Pull Request diffs, which creates a potential surface for indirect prompt injection attacks.
Ingestion points: External CI failure logs and PR diffs are fetched via scripts/ci-autofix/log-fetch.sh and injected into the manager-develop subagent prompt during the auto-fix phase.
Boundary markers: The instructions do not mandate the use of explicit boundary markers or 'ignore' instructions for the ingested log content.
Capability inventory: The skill has access to the Bash and Read tools, and interacts with the Git CLI to commit and push code changes.
Sanitization: There is no specified sanitization, escaping, or filtering of the CI log content before it is interpolated into the agent context.
[DYNAMIC_EXECUTION]: The skill dynamically generates code patches and applies them to the repository based on CI failures.
Evidence: The Phase 2 — Auto-Fix Loop describes a subagent proposing patches that are committed and pushed (fix(ci): auto-fix <classification> failure).
Context: While limited to mechanical fixes and 3 iterations, the skill facilitates the runtime generation of code that is subsequently executed within the CI pipeline.
[COMMAND_EXECUTION]: The skill relies on the execution of multiple shell scripts and the GitHub CLI to perform its core functions.
Evidence: The skill invokes scripts/ci-watch/run.sh, scripts/ci-autofix/log-fetch.sh, scripts/ci-autofix/classify.sh, and the gh pr checks command.
Protection: The skill explicitly prohibits high-risk commands such as git push --force, -f, or --force-with-lease.

hns-workflow-ci-loop